Codex Cloud: work while your laptop sleeps
Hand off a coding task and close your laptop. Here's how Codex Cloud environments, mobile access, code review, and plan limits shape the choice between cloud and local.
Codex Cloud can keep coding after your laptop goes to sleep. Prepare the repository, dependencies, tools, and access once, then start separate tasks from that reusable environment.
That's the useful part of the DevDay update. OpenAI's Cloud guide describes reviewing and continuing work across web, mobile, and desktop, so you can hand off a task and come back to it elsewhere.
A bug tied to your local simulator, an unfinished working tree, or a private service calls for a closer look before you hand it off. As of September 30, 2026, the choice comes down to what the task needs to run.
What Codex Cloud actually runs
A cloud environment is the reusable setup; a task is the work you start inside it. Each new task gets its own isolated workspace from the published environment, according to OpenAI's environment documentation.
Several tasks can use the same repository without agents editing the same working files. They share a starting point, while their changes and installed tools belong to their individual tasks.
Existing tasks keep their saved state. Republishing an environment changes the starting point for new tasks and leaves work already in progress alone. OpenAI warns that saved state is no substitute for source control, so commit work you need to preserve.
Think of the environment as a prepared development machine whose setup you can review. It needs the right repository, commands, credentials, and network access. Your local build working is a good start; you still need to check the cloud setup.
Codex Cloud vs local: hand off an issue or stay at the keyboard?
Local Codex works against your repository with tools installed on your machine, as the CLI documentation explains. Choose between that setup and Codex Cloud based on the work in front of you:
| Task or constraint | Sensible starting point | Reason |
|---|---|---|
| Focused change with clear acceptance criteria | Cloud | You can review a finished diff and test results later |
| Several independent investigations | Cloud | Separate task workspaces avoid a shared working tree |
| Debugging against local files or installed tools | Local | The necessary context is already on your machine |
| Rapid edits while you watch every command | Local | Feedback stays beside your editor and terminal |
| Work that should continue while the laptop sleeps | Cloud | The task's execution is remote |
| Repository setup that hasn't been tested remotely | Prepare the environment first | Dependency and access failures can stop useful work |
With the CLI, file edits and commands happen on your computer. That doesn't mean model inference happens there too, or that model requests stay on the device.
I'd use local Codex for debugging that keeps changing direction, and cloud for a bounded issue with a reproducible test. Before handing off a broad refactor, try a small task that uses the same build and service access.
Publish the setup before you start the task
OpenAI's getting started workflow begins in ChatGPT on the web or desktop:
- In a new task, choose Work in > Cloud and select or create an environment.
- Choose the GitHub repositories. Connect GitHub if prompted.
- Let Codex inspect the project, install dependencies and tools, and test the workflow.
- Supply missing information or access, then review the setup report and configuration.
- Save and publish the environment. Wait for the published confirmation before starting the task.
Installation finishing isn't the same as your project working. Check that a test of the project's behavior actually ran. If it needs a database, a private package registry, or a startup service, say so during preparation.
The environment docs distinguish direct environment variables from network secrets. Variables reach programs directly; for an allowed HTTPS destination, a network secret uses a placeholder that the proxy replaces with the credential. Shared setups can request personal values without sharing each person's credentials.
Configure and test the destinations your workflow needs. Allowing a domain doesn't supply a password or grant permission in that service. Before sharing a team setup, review both the prepared files and environment-owned credentials.
Give the agent an issue it can finish
Tell the agent what behavior to change, where to stop, and what evidence to return. An example prompt: fix the empty search results screen, preserve the existing filter default, and run the relevant UI tests.
Ask for the changed files, commands executed, test outcomes, and anything left unresolved. You can then assess the handoff without reading the entire transcript first.
Start on a phone, review on desktop
The new cloud environment docs explicitly support web, mobile, and desktop use. You can start an investigation on a phone and inspect the changes on a larger screen later. The work stays in the task's cloud workspace.
OpenAI's DevDay recap lists running on a computer, remotely from a phone, and in the cloud as separate possibilities. Using a phone as the interface doesn't tell you where the local tools run. Check that distinction when a workflow relies on your computer's tools.
A good phone task is easy to describe: investigate a failing test, explain a regression, or prepare a narrow fix. Save a large diff review for the desktop. You can start sooner from your phone, but you still need to understand the result.
Use /agents to track the work, then review the diff
The DevDay recap announces voice input for starting and steering CLI tasks, plus an /agents view for tracking delegated work. It lists the refreshed CLI and Code Review on all plans.
If you use the terminal, you need to see what each task is doing, which session needs input, and which result is ready. More concurrent work helps only while you can keep up with the reviews.
The Code Review docs cover pull request descriptions, diffs, comments, and checks. GitHub review is generally available; GitLab merge request support is in preview. Automatic cloud reviews can take an initial pass while you're away, as described in the recap.
Reviewing in chat doesn't approve or merge a change. Check each finding against the latest diff and expected behavior before sharing it. If a concern can't identify the triggering input or code path, it needs more investigation.
Codex Cloud pricing: access isn't unlimited usage
The launch recap lists Codex Cloud on Plus, Pro, Business, Healthcare, Education, and Enterprise. Those plans include access, with limits on usage.
OpenAI's pricing page lists Plus at $20 per month and Pro options at $100, $200, or $500 USD per month. Pro 500 includes Astra Ultrafast; ordinary cloud access doesn't require that top tier in the published plan description.
ChatGPT Work and Codex share usage. The pricing docs warn against using API token prices to estimate how many tasks a subscription includes. Look at task size, model choice, and your account's actual allowance before guessing how many jobs you can run each day.
API-key access has separate billing. OpenAI lists it for the CLI, SDK, and IDE extension at API rates, excluding cloud features such as GitHub code review. Buying API credits won't replace an eligible cloud plan.
A phone trial that hit trouble before the keynote
Simon Willison tried building his live-blog photo system with Codex Cloud on his phone on the way to DevDay. It ran into problems, so he switched to Claude Code for web. His first-hand live blog records a single trial of the earlier workflow, before the keynote's cloud announcement, rather than a test of the refreshed release.
In a Reddit discussion, EntertainmentSalt825 saw potential value in centralizing projects used across multiple systems. The same thread includes frustration over plan pricing. These individual reactions explain the appeal and the irritation; they don't measure reliability.
Try a small task in your own repository. Track the setup friction, whether tests ran, and how much review the result needed. That's what should decide whether you hand off the next issue.
FAQ
Does Codex Cloud work while my laptop is closed?
Yes. OpenAI says cloud tasks can continue while your computer sleeps. The environment needs the task's required access and tools.
Is Codex Cloud included in Plus?
Yes. Plus is an eligible cloud plan, subject to your account's usage allowance.
Is Codex Cloud better than local Codex?
Cloud suits bounded work you can hand off remotely. Choose local Codex when the task needs your machine's files, tools, or frequent interactive debugging.
Can Codex review GitLab merge requests?
Yes, in preview according to the current Code Review docs. GitHub review is generally available.